Academic
Research interests, education, and the systems they came out of.
I am finishing a Masters in Information Technology at Shahjalal University of Science and Technology, and I build applied AI systems that run in production — document pipelines and conversational agents in Bangla. I want to take that work into research.
I have no publications yet, so this page does not pretend otherwise. What it has instead is the specific problems I keep running into in deployed systems, and the systems themselves — which is the honest version of a research statement at this stage. My M.Sc. thesis is in progress; the topic will appear here once it is fixed.
Research interests
Document understanding and information extraction
I run a production pipeline that reads passports, green cards and birth certificates into structured records for an immigration and tax agency. The extraction itself turned out to be the easy half. The hard half is provenance and restraint: every value is stored as evidence attributed to the page it came from, a human can overrule any of it, and nothing is ever filed to a person automatically.
What I want to work on: extraction systems that know when they are wrong. Average field accuracy hides the only failures that matter, because the cost of a wrong date of birth is not the cost of a wrong middle initial.
NLP for low-resource and code-mixed language — Bangla and Banglish
I build WhatsApp ordering agents for shops in Bangladesh that handle Bangla, Banglish (romanised, code-mixed) and English — typed or sent as a voice note — and turn a rambling message into a structured order. Real customers do not write in the register the benchmarks are built from, and for the one they do write in there is very little labelled data and almost no evaluation to speak of.
What I want to work on: intent and slot extraction under code-mixing and transliteration variance, and speech in the accented, noisy, single-microphone conditions this actually runs in.
Applied LLM and agentic systems under production constraints
Reading one document costs about two cents. That number is a design constraint rather than a footnote, and so are latency, the failure modes of an over-confident model, and where a human has to sit in the loop. Systems that read untrusted documents also have an adversary the benchmark never had.
What I want to work on: evaluating agentic systems by what a human had to correct rather than by task success, and prompt-injection resistance in pipelines whose input is a stranger’s uploaded file.
Software systems for public institutions
A bilingual hall-management system for a university in Bangladesh, and before that EMIS — an education management platform with role-based access enforced at the API layer rather than hidden behind a UI toggle, built solo to production standards.
What I want to work on: what actually makes institutional software get adopted rather than abandoned — bilingual interfaces, tolerance for bad connectivity, and workflows shaped around how the office already works instead of how the schema does.
Systems behind them
LLM document pipeline for an immigration & tax agency
Schema-constrained extraction over real identity and financial documents, with human-reviewable evidence for every field. In production.
Bangla / Banglish WhatsApp ordering agents
Conversational order capture in a code-mixed low-resource language, text and voice, with human approval before anything is placed. Private client system — there is no repository to link and there never will be.
EMIS — education management platform
Django REST, React, PostgreSQL, JWT auth. Built alone to production standards; the commercial push did not work out. Source is public.
Education
- Masters in Information Technology in progress Shahjalal University of Science and Technology · expected 2026
- B.Sc. (Engg.) in Computer Science & Engineering Metropolitan University · 2023 · CGPA 3.66/4.00
- IELTS Academic Overall 7.5 · L 8.5 · R 7.0 · W 6.0 · S 7.5 (Dec 2023)
What I'm looking for
- PhD supervision — I am looking for a group working on applied language systems, and would rather start that conversation with a problem than a form.
- Funded M.S. positions and research assistantships.
- Research collaboration — if a group needs the systems half built and deployed rather than prototyped, that is the part I am already good at.
Available to start once the M.Sc. completes in 2026. Based in Sylhet, Bangladesh; willing to relocate.
Getting in touch
If any of the above overlaps with what your group works on, write to me — a paragraph about the problem is more useful to me than a formal enquiry, and you will get a straight answer about whether I am a fit.